Skip to content
focoit.com
Back to the canvas

Sequence

One question, from screen to audit trail

What happens to one question: gates, cache, retrieval, draft, verdict, review, delivery, and what is written to audit on the way.

One question, from screen to audit trail A sequence across six participants: site user, workflow, knowledge layer, model router, reviewer and audit. The workflow screens the question, logs that the gates passed and looks up the cache. On a hit it returns the reviewed answer with no model call. On a miss it retrieves claims-scoped grounding, drafts through the model router, applies the content filter and safety verdict, and sends a review request; the workflow idles while a checkpoint keeps the pending request across restarts. The reviewer accepts, edits, rejects or escalates; the full trail is audited and the answer is delivered or refused and cached. Site user Workflow code-defined graph Knowledge layer Model router Reviewer Audit 1 question + study + persona 2 PII screen → Prompt Shield 3 gates passed 4 cache lookup (SHA-256 key) CACHE HIT 5 reviewed answer · 0 LLM calls 6 retrieve (claims-scoped) 7 plan → subqueries → rerank 8 grounding + references 9 draft (Azure OpenAI · PHI path) 10 cited draft 11 content filter → safety verdict 12 review request (event) The workflow idles. A checkpoint keeps the pending request: it survives a restart or a deploy. 13 accept · edit · reject · escalate 14 full trail: question, spans, draft, verdict, action 15 deliver or refuse CACHE MISS LEGEND Call Return / audit Human review
Reference designOpen full size

What it shows

The same design as the overview, read in time. The workflow screens the question, records that the gates passed, and looks up the cache with a key built from the question, the knowledge version, the model version, the persona and the study.

On a hit the reviewed answer comes back with no model call. On a miss the workflow retrieves claims-scoped grounding, drafts through the model router, and applies the content filter and the safety verdict.

Then it sends a review request and idles. A checkpoint keeps the pending request, so a restart or a deploy in the middle of a review loses nothing. The reviewer accepts, edits, rejects or escalates; the full trail goes to audit, and the answer is delivered or refused.

This is a reference design for an agentic clinical-trials Q&A assistant. It is a design, not a description of a system built as shown.

Key decisions

  • Audit is written at the gates and again at the end, so a refused question leaves a record too.
  • The cache key includes the model and knowledge versions: changing either is a miss, with no manual invalidation.
  • The pause for review is a normal state of the workflow, not an error or a timeout.